Love the forensic approach to estimating MI430X's specs from Alice Recoque's config. The 211TF FP64 number feels spot-on given the ~70% HPL efficiency assumption. What really standouts is the HBM4 capacity advantage over Rubin (432GB vs 288GB), which could be huge for memory-hungry CFD simulations.
Your power numbers are all wrong here making the wrong assessments.
Typical power load for the 94 racks is 12MW.
TDP is 15MW.
Cooling is always factored up relative to TDP. No datacenter installs cooling at TDP.
Facilities power is also factored up to maximize power supply efficiency, generally 1.5x to 2x TDP.
That means the 2000W figure is a scaling factor for rack level TDP based on number of GPUs or something along those lines.
2500W could approximate per GPU power supply requirements.
15 MW for 94 racks means 160kW rack TDP.
The idea that cooling load = facility power = TDP is ridiculous.
Love the forensic approach to estimating MI430X's specs from Alice Recoque's config. The 211TF FP64 number feels spot-on given the ~70% HPL efficiency assumption. What really standouts is the HBM4 capacity advantage over Rubin (432GB vs 288GB), which could be huge for memory-hungry CFD simulations.
What an amazing piece of hardware. :)